A new study published in Machine Learning with Applications addresses this challenge with a reasoning-centric training methodology for multimodal large language models, or MLLMs. Led by Mohamed Bayan Kmainasi of Qatar Computing Research Institute and colleagues including Mucahid Kutlu, Ali Ezzat Shahroor, Abul Hasnat, and Firoj Alam, the work demonstrates that reinforcement learning with chain-of-thought supervision can push explainable meme moderation to state-of-the-art performance on both English hateful memes and Arabic propagandistic memes. The research represents, according to the authors, the first systematic study of group relative policy optimization, known as GRPO, in multimodal reasoning under cross-lingual, fine-grained, and self-training settings.

The team’s central insight is that thinking-based MLLMs, which generate explicit intermediate reasoning steps before committing to an answer, are ideally suited to memes because their meaning depends on image-text interaction rather than unimodal cues. But whether the reasoning capabilities of such models, typically honed on mathematics and code generation, transfer to subjective, culturally situated tasks like meme moderation remained an open question. Standard supervised fine-tuning alone provides limited control over the balance between prediction correctness and rationale faithfulness, since cross-entropy loss treats all output tokens equally regardless of their functional role.

To resolve this, the researchers designed a three-stage training pipeline. The first stage is a supervised fine-tuning warm-up that aligns the model with gold labels, natural language explanations, and distilled reasoning traces produced by GPT-4.1. The second stage applies GRPO with a composite reward function that jointly optimizes classification correctness, output-format compliance, reference-based explanation similarity measured by METEOR, explanation length, and a novel thinking-length reward called Rthink. The third stage, self-training GRPO or ST-GRPO, extends the approach to unlabeled data using consensus-based pseudo-labels derived from the model’s own majority-vote predictions.

The composite reward is carefully weighted so that the two binary objectives, label correctness and format compliance, each receive 0.35 and together dominate the three auxiliary components, which contribute 0.30 combined. The thinking-length reward is particularly significant. Without it, the authors observed a consistent reward-hacking pattern: the model learned to compress or empty its reasoning traces while still collecting high reward, a length-shortening bias that was especially pronounced on the harder Arabic dataset. Rthink penalizes only reasoning traces shorter than a minimum threshold of 150 words, discouraging degenerate outputs without incentivizing verbosity.

The evaluation spans two distinct tasks and two languages. The English benchmark is the Facebook Hateful Memes dataset, containing roughly 11,000 memes where classification requires joint multimodal understanding. The Arabic benchmark, ArMeme, contains about 5,700 memes with four labels covering propaganda, not-propaganda, not-meme, and other. Because no fine-grained propaganda annotations existed for ArMeme, the team built a dual-annotator pipeline using GPT-4.1 and Llama-4-Scout as independent labelers of 23 propaganda techniques, consolidated by Gemini-3-Pro as an arbiter. Human validation on 584 memes showed the consolidated annotations aligned better with human reference labels than either single-model source.

The results are striking. On the Hateful Memes benchmark, the best supervised GRPO configuration with thinking-length regularization achieved 82.0 percent accuracy and 0.80 macro-F1, outperforming prior reported results and strong sequence-classification baselines such as Qwen3-VL-8B-Instruct and Gemma-3-12B-IT. On ArMeme, self-training GRPO reached 0.612 macro-F1, improving over previous work by 7.6 points and over the original ArMeme benchmark by 6.1 points. Notably, unimodal baselines lagged far behind: the best text-only model on ArMeme reached only 0.509 macro-F1, and image-only models averaged just 0.267, confirming that cross-modal reasoning captures signals that neither modality provides alone.

The self-training stage showed a telling asymmetry. On ArMeme, where unlabeled data was collected from the same social media sources as the labeled set, ST-GRPO improved macro-F1 by 1.5 points with gains concentrated in minority classes. On the English dataset, it slightly degraded performance, a result the authors attribute to distribution mismatch between the unlabeled pool and the benchmark, and to majority-vote bias amplification in binary classification, where small prediction biases can produce near-unanimous consensus that reinforces rather than corrects class skew. The finding suggests that consensus-based pseudo-labeling requires both distribution alignment and sufficient label-space diversity to provide reliable supervision.

Beyond raw accuracy, the model produces natural language explanations alongside its predictions, a property that sequence classifiers lack. Modality ablations showed the model genuinely depends on both inputs, with image removal hurting the English task most and OCR text removal crippling Arabic propaganda detection. An LLM-as-judge evaluation using GPT-4.1 and Gemini-2.5-Pro found the trained models substantially outperformed the zero-shot backbone on grounding, correctness, and usefulness for moderation, approaching the quality of human-verified reference explanations. The team has released all code, data extensions, prompting templates, and evaluation resources on GitHub.

The implications extend well beyond memes. The study shows that fine-grained supervision and distilled chain-of-thought rationales complement reinforcement-learning-based optimization, that multi-LLM annotation pipelines can scale fine-grained labeling to previously unlabeled domains, and that even a simple thresholded reasoning-length penalty can stabilize RL training against reward hacking in multimodal settings. For content moderation at scale, where subjective interpretation of culturally embedded imagery is the daily reality, the work offers a concrete template for building systems that not only flag harmful content but also articulate why, enabling meaningful human review rather than opaque automated verdicts.

Subject of Research: Reinforcement learning with chain-of-thought supervision for explainable detection of hateful and propagandistic memes

Article Title: Adapting reinforcement learning with chain-of-thought supervision for explainable detection of hateful and propagandistic memes

Article References: Kmainasi, M. B., Kutlu, M., Shahroor, A. E., Hasnat, A., & Alam, F. (2026). Adapting reinforcement learning with chain-of-thought supervision for explainable detection of hateful and propagandistic memes. Machine Learning with Applications, 26, Article 101003. https://doi.org/10.1016/j.mlwa.2026.101003

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.101003

Keywords: reinforcement learning, chain-of-thought, multimodal large language models, hateful memes, propaganda detection, content moderation, GRPO, self-training, explainable AI, Arabic memes, reward hacking, fine-grained annotation

The core challenge the researchers set out to solve is one that plagues mineral exploration worldwide. Stream sediment geochemical surveys, which measure the concentrations of dozens of elements in sediment samples collected from drainage networks, produce vast datasets. Yet only a tiny fraction of the sampled locations can be confidently labeled as mineralized, because confirmed deposits are rare and expensive to verify. Standard supervised classifiers, which learn from labeled examples, struggle when the positive class is so sparsely represented. Meanwhile, the overwhelming majority of samples remain unlabeled, a reservoir of information that conventional methods simply ignore.

Zheng and Chen’s framework attacks this problem from three directions at once. First, LightGBM serves as the base classifier, chosen for its ability to capture the nonlinear relationships between element concentrations and the spatial distribution pattern of mineral deposits. LightGBM is a gradient boosting decision tree algorithm known for its speed and memory efficiency, achieved through techniques such as histogram-based splitting and leaf-wise tree growth. In mineral exploration, where the interplay between trace element signatures and ore-forming processes is anything but linear, this flexibility matters enormously.

Second, the framework employs a self-training algorithm, a semi-supervised technique in which a model is initially trained on the small labeled set and then iteratively predicts labels for the unlabeled data. The most confident predictions are added to the training pool, and the model is retrained, gradually bootstrapping its way toward a more complete understanding of the data. Self-training has a long pedigree in machine learning, dating back to the 1960s, but its application to geochemical anomaly recognition is relatively recent. By exploiting the vast unlabeled portion of stream sediment data, the method effectively converts a data-poor problem into a data-rich one.

Third, the researchers introduced the Synthetic Minority Oversampling Technique, or SMOTE, to address the severe class imbalance that would otherwise cripple the initial training stage. Because mineralized locations are so rare, a classifier trained naively would learn to simply predict that nothing is mineralized, achieving high accuracy while being useless in practice. SMOTE works by generating synthetic examples of the minority class, interpolating between existing positive samples in feature space rather than merely duplicating them. This enlarges the number of positive samples enough for LightGBM to establish a meaningful model during the first round of self-training, after which the iterative process takes over.

To test the framework, the team built four models on interpolated 1:50,000-scale stream sediment geochemical data from Moridawa: the self-training LightGBM, a self-training support vector classifier, a standalone LightGBM, and a standalone support vector classifier. This head-to-head comparison allowed the researchers to isolate the contributions of both the self-training strategy and the choice of base algorithm. The results were clear. Among the four models, the self-training LightGBM achieved the most favorable overall balance between classification performance, spatial prediction efficiency, and computational efficiency, outperforming its rivals across the evaluation metrics used in the study.

The spatial results are perhaps the most striking. The geochemical anomalies recognized by the self-training LightGBM model cover only 2.74 percent of the entire study area, yet they encompass all of the known molybdenum deposits. In exploration terms, this is exactly what one wants: a small, high-confidence footprint that directs drilling and follow-up fieldwork toward the most promising ground. Anomalies that blanket huge swaths of terrain may technically capture the deposits, but they offer little practical guidance and inflate exploration costs.

Equally important is how the predicted anomalies align with the region’s geology. The team found close spatial consistency between the recognized anomalies and major faults, Mesozoic intermediate to acidic intrusions, and the contact zones where those intrusions meet the surrounding country rock. This makes geological sense, since such intrusion-country-rock contact zones are classic loci for ore-forming fluids, and faults often serve as conduits for mineralizing fluids. The fact that the model’s predictions converge on these structurally and magmatically favorable settings suggests that the algorithm is genuinely learning the fingerprints of mineralization rather than fitting statistical noise.

The implications extend beyond a single case study in Inner Mongolia. Mineral exploration increasingly relies on machine learning to sift through ever-larger geochemical and geophysical datasets, but the field has been hampered by the same recurring obstacles: complex geological settings, extreme class imbalance, and sparse labels. By combining a fast, nonlinear base learner with semi-supervised self-training and targeted oversampling, the new framework offers a template that could be adapted to other deposit types, other elements, and other survey scales. The authors note that the method provides a useful application approach for recognizing mineralization-caused anomalies from sparse labeled and vast unlabeled geochemical data collected in complex geological settings.

There are, of course, caveats. The framework was validated in one area, and its performance elsewhere will depend on data quality, the representativeness of the labeled samples, and the specifics of local geology. The researchers also acknowledge that no datasets were generated or analyzed beyond those used in the case study, meaning broader benchmarking remains future work. Still, the study, funded by the National Natural Science Foundation of China under grant number 42472361, represents a meaningful step toward smarter exploration. As the global demand for critical metals such as molybdenum continues to climb, tools that can squeeze more predictive power out of existing survey data, without demanding expensive new field campaigns, are likely to find an eager audience across the mining industry and academic geochemistry alike.

Subject of Research: A self-training LightGBM machine learning framework for recognizing mineralization-caused geochemical anomalies in stream sediment data.

Article Title: A self-training framework based on LightGBM for recognizing mineralization-caused geochemical anomalies

Article References: A self-training framework based on LightGBM for recognizing mineralization-caused geochemical anomalies. (n.d.). https://doi.org/10.1007/s12145-026-02239-y

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02239-y

Keywords: LightGBM, self-training, geochemical anomalies, mineral exploration, SMOTE, semi-supervised learning, molybdenum deposits, Inner Mongolia, stream sediment data, machine learning, mineral prospectivity, Jilin University